Western pacific ocean subelevation ridge line forecasting method based on abnormal weighting
By combining weighted diagonal averaging and self-attention mechanism, the problem of insufficient feature utilization in the forecasting of the western Pacific subtropical high ridge is solved, improving the forecast accuracy and robustness. It is applicable to the forecasting of the western Pacific subtropical high ridge in the field of meteorological forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies fail to fully utilize the periodic and residual characteristics of RI sequences in forecasting the western Pacific subtropical high ridge, and placing the self-attention mechanism at the LSTM input may destroy the feature structure, resulting in insufficient prediction accuracy.
We employ a weighted diagonal averaging strategy and a self-attention mechanism to embed the hidden state layer of an LSTM based on the features of anomalous years. We process the SSA component through differential weights and integrate the physically meaningful ridge index component into the LSTM network to construct a deep-shallow feature fusion enhancement module.
It improves the prediction accuracy of RI sequences, enhances the ability to extract key information, alleviates the memory decay problem of LSTM, and achieves more accurate forecasts of the western Pacific subtropical high ridge.
Smart Images
Figure CN121679751A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological forecasting, specifically to a forecasting method for the western Pacific subtropical high ridge based on anomaly weighting. Background Technology
[0002] The Western Pacific subtropical high is an important circulation system that triggers extreme heat and rainstorms. Changes in its ridge index (RI) indicate the north-south movement of the Western Pacific subtropical high. Accurate prediction of the RI can indicate the risks of high temperatures, rainstorms, or typhoons that relevant regions may face in the future, which is of great significance for scientific emergency response and reducing disaster losses.
[0003] Predicting the ridgeline of the Western Pacific subtropical high based on stacked long short-term memory (LSTM) networks is currently one of the mainstream techniques. Regarding RI prediction, deep learning-based methods can automatically extract complex nonlinear relationships between features without relying on manual experience to define feature associations, and are therefore gaining increasing attention. In particular, the use of singular spectral analysis (SSA), a signal processing technique, effectively reconstructs features from the original complex time series, exhibiting stronger adaptability compared to traditional wavelet transform or empirical mode decomposition methods.
[0004] While the prediction method combining SSA and LSTM has demonstrated powerful predictive capabilities, it still has two limitations: First, SSA employs a uniform processing strategy for all decomposed components, applying an indiscriminate diagonal averaging operation to all components during the reconstruction stage. This fails to fully consider the different characteristics of the periodic and residual terms in the RI sequence. Specifically, the RI sequence exhibits significant periodicity, and the residual terms also contain crucial information with predictive value. Applying the same processing method to all components makes it difficult to fully utilize the regularity features in the periodic components and effectively capture anomalous information in the residual terms. Second, existing methods often place the attention mechanism at the LSTM input for feature fusion. Because this requires secondary weighting of the enhanced features, it may disrupt the feature structure extracted by SSA. Summary of the Invention
[0005] To address the aforementioned problems, this invention proposes a method for predicting the ridge of the Western Pacific Subtropical High based on anomaly weighting. A weighted diagonal averaging strategy tailored to the characteristics of anomalous years is designed to maintain the periodicity of the sequence and enhance the extraction of key information. Simultaneously, a self-attention mechanism is embedded into the hidden state layer of the LSTM, enabling the integration of the LSTM network learning results with the physically meaningful ridge index components, thus promoting the effective coupling of data-driven knowledge and physical prior information. The technical solution provided by this invention is as follows:
[0006] A method for forecasting the ridge of the Western Pacific subtropical high based on anomaly weighting is proposed, comprising data preparation, data processing, and result output. The data processing includes two modules: a weighted diagonal averaging module for anomalous year characteristics and a deep-shallow feature fusion enhancement module. First, the original ridge index sequence obtained in the data preparation stage is used as the input to the weighted diagonal averaging module. The main period is determined through power spectrum analysis, and using this as a time window, the original ridge index sequence is decomposed into 6 subsequences through singular spectrum analysis. Each subsequence includes 1 trend component, 2 periodic components, and 3 residual components. The periodic component and residual component are differentially weighted in combination with anomalous year information to construct a new feature matrix, which is then input into the deep-shallow feature fusion enhancement module. Finally, the predicted ridge index is obtained.
[0007] The preferred processing flow for the weighted diagonal averaging module is as follows:
[0008] S11. The original ridge index sequence is analyzed by Fourier transform and power spectrum. By identifying the significant peaks in the power spectrum, the main period of the sequence is determined and used as a time window.
[0009] S12, based on a defined time window, uses singular spectral analysis to decompose the original sequence into 6 sub-components;
[0010] S13, Based on the Z-score method, identify the set of abnormal years, and perform differentiated weight adjustments on the periodic components and residual components of abnormal years. The weight of abnormal years in the periodic component is reduced by 20%, and the weight of abnormal years in the residual component is increased by 50%.
[0011] S14 integrates all processed sub-components into a new feature matrix as input to the deep-shallow feature fusion enhancement module.
[0012] Preferably, power spectrum The formula is as follows:
[0013]
[0014] in, Angular frequency, For the sample size, It is the functional expression of the signal in the time domain. It is the imaginary unit.
[0015] Preferably, the 6 sub-components are ,in For time points, trend items The sequence represents the long-term evolutionary background, reflecting the changing trend of the north-south position of the subtropical high ridge; periodic term and Represents regular periodic fluctuations in the sequence, where The corresponding main cycle signal reflects the most stable intraseasonal oscillation of the subtropical high. The corresponding secondary period signal serves as a supplement to the primary period; residual term , and This represents irregular high-frequency fluctuations in the sequence, where Includes short-term fluctuations with some correlation. and It is a weak residual signal, whose fluctuations tend to be more random and whose signal strength decreases sequentially.
[0016] Preferably, the Z-score method identifies the set of anomalous years. The threshold is 2.0, and the formula is as follows:
[0017]
[0018] in and It is the original sequence The mean and standard deviation, To represent a time series at time points The observed values.
[0019] The preferred deep-shallow feature fusion enhancement module processing flow is as follows:
[0020] S21, at each time step, the new feature matrix constructed by the weighted diagonal averaging module is used as the input feature. Hidden states are generated sequentially through LSTM units. ;
[0021] S22, hide state Input is weighted and fused using self-attention, integrating global contextual information to output a new hidden state. ;
[0022] S23, All Hidden Status As input to the fully connected layer, the ridge index prediction is output after being mapped by the fully connected layer.
[0023] The preferred formula for the self-attention mechanism is as follows:
[0024]
[0025] in , and These are the query weight matrix, key weight matrix, and value weight matrix, respectively. Scaling factor It is the attention dimension; and The original features are the input. This is a new feature for the output.
[0026] Compared with the prior art, the beneficial effects achieved by the present invention are:
[0027] This invention employs a weighted diagonal averaging strategy, using the Z-score method to detect low-value and high-value anomalous years, and then assigning differentiated weights to data points of different years in the periodic and residual terms. Finally, weighted reconstruction is achieved during the diagonal averaging process, which can effectively preserve the key information of anomalous years in the SSA subsequence, while reducing sequence complexity and noise interference, thereby obtaining more accurate intrinsic dynamic characteristics of the RI sequence.
[0028] This invention constructs a deep-shallow feature fusion enhancement module, introducing a self-attention mechanism into the hidden state layer of an LSTM. The hidden state sequence output by the LSTM is used as the input to the self-attention mechanism. By calculating the correlation between the hidden state sequences, dynamic weighting of the hidden outputs at different time steps is achieved. FSAEM fully utilizes historical context information during the modeling process, helping to alleviate the memory decay problem of LSTM during iterative iterations. Simultaneously, the introduction of the Self-Attention mechanism enhances the model's ability to filter and integrate historical states, thus better aligning with the temporal changes of the RI (Regression Analysis). Attached Figure Description
[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0030] Figure 1 This is an overall flowchart of the present invention;
[0031] Figure 2 This is a flowchart of the deep-shallow feature fusion enhancement module of the present invention;
[0032] Figure 3 It is the result of dividing the test set samples in the comparative experiment under different RI change modes;
[0033] Figure 4 It is the prediction accuracy index of each model in three types of samples in the comparative experiment under different RI change modes. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] To make the above-mentioned objectives, features and effects of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0036] Example 1: A method for forecasting the western Pacific subtropical high ridge based on anomaly weighting, including data preparation, data processing, and result output. Data processing comprises two modules: a weighted diagonal averaging module for anomalous year characteristics and a deep-shallow feature fusion and enhancement module. Its workflow is as follows: Figure 1 As shown, the original ridge index sequence is first used as input to the weighted diagonal averaging module. The main period is determined by power spectrum analysis, and then the original ridge index sequence is decomposed into 6 subsequences by singular spectrum analysis using this as a window. Based on this, the period and residual components are differentially weighted in combination with anomalous year information to construct a new feature matrix, which is then input to the deep-shallow feature fusion enhancement module. Finally, the prediction of the ridge index is obtained.
[0037] Specifically, the weighted diagonal averaging module processing flow is as follows:
[0038] S11, the original ridge index sequence is analyzed using Fourier transform and power spectrum, as shown in the following formula:
[0039]
[0040] in, Angular frequency, For the sample size, It is the functional expression of the signal in the time domain. It is the imaginary unit, which is used to identify the power spectrum. The significant spectral peaks were identified, the main period T of the sequence was determined, and T was used as the unified time window length for subsequent singular spectral analysis decomposition and LSTM network.
[0041] S12, based on a defined time window ( Using singular spectral analysis to analyze the original sequence Decomposed into 6 sub-components ,Right now:
[0042]
[0043] in, For time points, trend items The sequence represents the long-term evolutionary background, reflecting the changing trend of the north-south position of the subtropical high ridge; periodic term and Represents regular periodic fluctuations in the sequence, where The corresponding main cycle signal reflects the most stable intraseasonal oscillation of the subtropical high. The corresponding secondary period signal serves as a supplement to the primary period; residual term , and This represents irregular high-frequency fluctuations in the sequence, where Includes short-term fluctuations with some correlation. and It is a weak residual signal, whose fluctuations tend to be more random and whose signal strength decreases sequentially.
[0044] S13, Identifying sets of abnormal years based on the Z-score method Based on the physical characteristics of the components, differentiated weight adjustments are made to specific components of abnormal years. As an example, this embodiment identifies a set of abnormal years. The threshold is 2.0:
[0045]
[0046] in and It is the original sequence The mean and standard deviation, To represent a time series at time points The observed values.
[0047] For periodic components and In abnormal years, the weight is reduced by 20% to maintain the stability of the cyclical pattern, that is:
[0048]
[0049] For residual components , and The weight of abnormal years is increased by 50% to amplify the abnormal signal, that is:
[0050]
[0051] in These are the original components of the SSA decomposition. This is the component adjusted after the abnormal year.
[0052] S14, process all processed components The features are integrated into a new feature matrix and used as input to the deep-shallow feature fusion enhancement module.
[0053] Specifically, the deep-shallow feature fusion enhancement module, such as Figure 2 As shown, the processing flow is as follows:
[0054] S21, at each time step, the new feature matrix constructed by the weighted diagonal averaging module is used as the input feature. Hidden states are generated sequentially through LSTM units. ;
[0055] S22, hide state Input is weighted and fused using self-attention, integrating global contextual information to output a new hidden state. The formula is as follows:
[0056]
[0057] in , and These are the query weight matrix, key weight matrix, and value weight matrix, respectively. Scaling factor It is the attention dimension; and The original features are the input. This is a new feature for the output.
[0058] S23, All Hidden Status As input to the fully connected layer, the ridge index prediction is output after being mapped by the fully connected layer.
[0059] Example 2: Experimental analysis of the method of this invention was conducted using 500 hPa monthly mean geopotential height field data from the NCEP / NCAR reanalysis dataset. This dataset, a widely used benchmark in climate research, integrates multi-source historical observation data and numerical model outputs based on an advanced data assimilation system, providing a global gridded atmospheric variable series with a spatial resolution of 2.5° × 2.5° since 1948, exhibiting temporal continuity and spatial consistency. Before inputting the time-series data into the model of this invention, the corresponding ridge index series must first be calculated monthly based on the geopotential height field according to the physical definition of RI, in order to construct the standardized input required by the model. The RI dataset used covers May–October of each year from 1948 to 2024, containing a total of 462 valid samples. To further evaluate model performance, the complete sequence was divided into three independent subsets in chronological order: the training set (1948–1994, accounting for 60% of the total sample) was used to learn the natural variability features of RI; the validation set (1995–2009, accounting for 20%) was used to assist in hyperparameter tuning through an early stopping mechanism; and the test set (2010–2024, accounting for 20%) was used to independently evaluate the model's generalization ability and predictive stability.
[0060] Three metrics—R² (fit coefficient), MSE (mean squared error), and MRE (mean relative error)—were used to evaluate the model's predictive performance. These three metrics reveal the model's performance characteristics from different dimensions, complementing each other and providing a comprehensive and objective measure of the model's predictive ability. The calculation formulas for these metrics are as follows:
[0061]
[0062]
[0063]
[0064] in These are actual observed values. For predicted values, This is the average of the actual observed values. This refers to the sample size. MSE and MRE quantify the accuracy of the prediction by calculating the difference between the model's predicted values and the actual observed values. R² measures the model's ability to explain the overall fluctuations in the data and is a core indicator for evaluating goodness of fit. These three metrics are widely used performance evaluation standards in the field of predictive modeling. Lower MSE and MRE, and higher R² indicate that the model has excellent predictive performance and generalization ability.
[0065] Experiment 1: Three representative time series prediction methods were selected for comparative analysis, including LSTM, iTransformer, and SSA-LSTM. iTransformer employs an inverted structure design, treating the entire time series of each variable as an independent label, and encodes deep features in the time dimension through a feedforward network. SSA-LSTM combines the advantages of singular spectrum analysis and deep learning methods. First, singular spectrum analysis decomposes the original sequence into trend, periodic, and residual components. Signal reconstruction effectively improves data quality. Then, an LSTM network learns more robust temporal features from the cleaned data, significantly improving prediction accuracy. All experiments were run on a Windows 11 operating system environment, implemented using Python 3.9 and PyTorch 2.7.1 deep learning frameworks to ensure high consistency of the experimental environment and reproducibility of results. All computational tasks were accelerated using an NVIDIA GeForce RTX 5070Ti GPU. In the comparative experiments, the hyperparameters of all networks were automatically tuned using a Bayesian optimizer on the validation set, with 20 iterations to balance search efficiency and parameter quality.
[0066] Based on the test set, RI predictions were performed on the proposed method and the comparative method over a period of 6 months. The quantitative evaluation results are shown in Table 1.
[0067] Table 1. Accuracy Evaluation Results of the NCEP / NCAR Dataset
[0068]
[0069] As shown in the table above, this invention achieves a mean squared error (MSE) of 2.97, a fitting coefficient (R²) of 0.83, and a mean relative error (MRE) of approximately 4.16% in RI predictions over six consecutive months, outperforming other comparative models. Traditional LSTM models, when processing raw RI sequences, are often limited by the non-stationarity and complex nonlinear patterns within the data, making it difficult to capture their deep-seated patterns and thus affecting prediction accuracy. Although new prediction models such as iTransformer and SSA-LSTM have been proposed in recent years, their performance improvement is limited due to their failure to fully incorporate the unique temporal structure of RI sequences. FSALSTM, by introducing TAWDA and FSAEM, achieves targeted temporal modeling at the structural level, thereby improving prediction accuracy. Overall, FSALSTM demonstrates good accuracy and can provide effective support for RI prediction.
[0070] Experiment 2: To verify the effectiveness of the proposed weighted diagonal averaging module, an ablation experiment was designed for comparison: the complete model of this invention was used as the experimental group; the model after removing the weighted diagonal averaging module was used as the control group, that is, after SSA decomposition, multiple sub-sequences were directly input into the deep-shallow feature fusion enhancement module.
[0071] Table 2 Comparison of ablation experimental results with removal of weighted diagonal averaging module
[0072]
[0073] The accuracy evaluation results show that the experimental group outperformed the control group in all accuracy metrics for RI prediction. Specifically, the MRE decreased from 4.19% to 4.16%, the MSE decreased from 3.22 to 2.97, and the R² increased from 0.81 to 0.83. The fundamental reason for this performance improvement is that the weighted diagonal averaging module effectively suppressed the interference of stationary background noise in the time series, allowing the feature sequences reconstructed from the singular spectrum to more clearly highlight key anomalous patterns related to RI events. This optimization provides the LSTM model with more discriminative and higher signal-to-noise ratio dynamic features during the preprocessing stage, thereby significantly enhancing its ability to capture complex nonlinear time series evolution patterns and improving prediction accuracy.
[0074] Experiment 3: To separately verify the effectiveness of the designed deep-shallow feature fusion enhancement module, another ablation experiment was designed: the complete model of this invention was used as the experimental group; the model after removing the deep-shallow feature fusion enhancement module was used as the control group for comparison, reducing the model's ability to dynamically weight the entire sequence context. By comparing the performance of the two groups of models, the actual contribution of the deep-shallow feature fusion enhancement module was quantified.
[0075] Table 3 Comparison of ablation experimental results for removing deep-shallow feature fusion enhancement modules
[0076]
[0077] As shown in the table above, the accuracy evaluation results demonstrate that the experimental group outperformed the control group in all accuracy metrics for RI prediction. Specifically, MRE decreased from 4.51% to 4.16%, MSE decreased from 3.61 to 2.97, and R² increased from 0.79 to 0.83. Due to a lack of global coordination, the memory carried by the hidden states of LSTM is often localized and discontinuous, making it difficult for the model to capture complete sequence patterns. The deep-shallow feature fusion enhancement module, by introducing a self-attention mechanism, strengthens the model's retention and utilization of historical information, resulting in more globally consistent predictions.
[0078] Experiment 4: Based on the Z-score, the test set samples were divided into three typical RI change patterns: stationary period (|Z|≤1), fluctuating period (1<|Z|≤2), and abnormal period (|Z|>2). By analyzing the prediction performance of the method of this invention and the comparative model under the three change patterns, it was verified whether FSALSTM can maintain stable and reliable prediction performance in different stages of RI sequence change. Figure 3 The results of the test set sample partitioning are shown; Figure 4 The prediction accuracy metrics of various models were compared across three types of samples: stable period, fluctuating period, and abnormal period, including MSE and MRE.
[0079] like Figure 3 As shown, the variation pattern of RI during the testing period can be divided into three stages: a stable period (accounting for about 67.5%), during which the main body of the subtropical high is in a stable position with little movement; a fluctuating period (accounting for about 27.6%), characterized by the north-south swing of the ridge line, which often causes the rain belt to move and the range of high temperatures to change; and an abnormal period (accounting for about 4.9%), specifically manifested as the "southward retreat" in August 2012 and July-August 2018, and the "northward shift" in May 2023.
[0080] like Figure 4 As shown, in terms of model performance, the prediction metrics of this invention are superior to the comparative model under all variation modes. Specifically, during the stable period, the MSE of this invention is 2.5 and the MRE is 3.9%; during the fluctuating period, the MSE is 2.9 and the MRE is 6.1%; and during the abnormal period, the MSE is 9.92 and the MRE is 9.6%. In terms of the improvement in prediction accuracy, the MSE of this invention is reduced by about 1.3 compared to the comparative model during the stable period; and the MRE is reduced by more than 6% during the abnormal period. This result verifies that the synergistic effect of the weighted diagonal averaging module and the deep-shallow feature fusion enhancement module of this invention can maintain stable prediction performance under different RI variation modes, demonstrating the robustness of the model.
[0081] Example 3: The computer-readable storage medium of this example stores a computer program that, when executed by a processor, implements the steps in the method for predicting the western Pacific subtropical high ridge based on anomaly weighting in Example 1.
[0082] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.
[0083] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.
[0084] Example 4: The computer device of this example includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for predicting the western Pacific subtropical high ridge based on anomaly weighting in Example 1.
[0085] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.
[0086] Those skilled in the art will clearly understand that each implementation can be achieved using software plus the necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for forecasting the ridge line of the western Pacific subtropical high based on abnormal weighting, comprising data preparation, data processing and result output, characterized in that, The data processing includes two modules: a weighted diagonal average module for characteristics of abnormal years and a deep-shallow layer feature fusion reinforcement module. First, the original ridge line index sequence obtained in the data preparation stage is taken as the input of the weighted diagonal average module. The main period is determined through power spectrum analysis, and the original ridge line index sequence is decomposed into six sub-sequences through singular spectrum analysis with the main period as the time window. The sub-sequences include one trend component, two periodic components and three residual components. The periodic components and residual components are differentially weighted in combination with abnormal year information to construct a new feature matrix input into the deep-shallow layer feature fusion reinforcement module, and finally the prediction of the ridge line index is obtained.
2. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 1, characterized in that, The processing procedure of the weighted diagonal average module is as follows: S11, the Fourier transform and power spectrum analysis are performed on the original ridge line index sequence. The main period of the sequence is determined by identifying the significant spectral peak of the power spectrum, and is taken as the time window; S12, based on the determined time window, the original sequence is decomposed into six sub-components by singular spectrum analysis; S13, the abnormal years are identified based on the Z-score method. The periodic components and residual components of the abnormal years are differentially weighted. The abnormal year weight of the periodic component is reduced by 20%, and the abnormal year weight of the residual component is increased by 50%; S14, all the processed sub-components are set into a new feature matrix as the input of the deep-shallow layer feature fusion reinforcement module.
3. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 2, characterized in that, Power spectrum The formula is as follows: ； wherein is the angular frequency, is the number of samples, is the function expression of the signal in the time domain, is the imaginary unit.
4. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 2, characterized in that, 6 sub-components are where is the time point, the trend term represents the evolution background of the sequence long-term, reflecting the trend of the north-south position of the subtropical ridge line; the periodic term and represent the regular periodic fluctuations in the sequence, where corresponds to the main cycle signal, reflecting the most stable seasonal intra-seasonal oscillation of the subtropical high, corresponds to the secondary cycle signal, which is a supplement to the main cycle; the residual term , and represent irregular high-frequency fluctuations in the sequence, where contains short-term fluctuations with certain correlation, and are weak residual signals, whose fluctuations tend to be random and the signal strength decreases in turn.
5. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 4, characterized in that, Z-score method identifies the set of abnormal years with a threshold of 2.0, as follows: ； wherein and is the mean and standard deviation of the original sequence is the observation of the time series at time point . 6. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 1, characterized in that, The processing procedure of the deep-shallow layer feature fusion reinforcement module is as follows: S21, in each time step, the weighted diagonal average module constructs a new feature matrix as input features The hidden state is generated sequentially via the LSTM unit ; S22, the hidden state The input self-attention is weighted and fused, and global context information is integrated to output a new hidden state ; S23, all hidden states As input to the fully connected layer, the final ridge index prediction value is output after mapping via the fully connected layer.
7. The abnormal weighting based method for forecasting the ridge line of the western Pacific subtropical high according to claim 6, characterized in that, The self-attention mechanism formula is as follows: ； wherein , and are the query weight matrix, key weight matrix and value weight matrix, respectively; is a scaling factor, is the attention dimension; and are the original features of the input, is the new feature of the output.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the abnormal weighting based western pacific subtropical high ridge line prediction method according to any one of claims 1-7.
9. A computer device comprising a processor, a memory and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the abnormal weighting based western pacific subtropical high ridge line prediction method according to any one of claims 1-7.